"Frontier": A Word That's Traveling Too Fast

This month, I heard half a dozen organizations introduce themselves as "frontier". None of them trains a model. Maybe that's fine. Maybe it isn't. Either way, it deserves a closer look.

---

The Word Is Everywhere

On LinkedIn, in pitch decks, in Microsoft press releases, in C-level presentations: frontier. Frontier model, frontier company, frontier firm, frontier-ready, frontier-grade. In eighteen months, the term has migrated from AI-governance research circles to a buzzword as ubiquitous as "cloud-native" was ten years ago.

The problem is that it no longer means quite the same thing depending on who's saying it. And like every word that drifts fast, it becomes both irresistible and risky.

Where the Word Comes From

Originally — meaning 2023 — the term frontier model came out of AI-governance research. The idea was precise: name the general-purpose AI models that push the state of the art far enough to pose systemic risks (cybersecurity, disinformation, weapons proliferation, loss of control).

The definition isn't vague:

• The Frontier Model Forum (Anthropic, OpenAI, Google, Microsoft) talks about models trained with a compute budget on the order of 10^26 FLOPS.
• The EU AI Act uses a 10^25 FLOPS threshold to qualify "high-impact" models.
• The defunct California SB 1047 used similar thresholds.

It's a technical threshold, not a marketing one. We're talking about training runs that cost a hundred million dollars and up, that mobilize dedicated GPU farms, and that are by nature out of reach for 99.99% of the organizations on the planet.

Frontier, in its original meaning, doesn't describe an attitude. It describes an electricity bill.

A Very Small Club

Direct consequence: frontier AI companies, in the strict sense, form a very small club. Anthropic, OpenAI, Google DeepMind, xAI, Meta AI, and a handful of Chinese labs (DeepSeek, Alibaba Qwen, Moonshot). That's about it.

They aren't the organizations that use the best models. They're the ones that train them. The distinction isn't cosmetic: it maps to emerging regulatory obligations, to budgets that have nothing in common with those of an SMB, and to a quasi-geopolitical status.

A company that consumes the Anthropic API, even in very sophisticated ways, isn't frontier in Anthropic's sense. It's a customer.

The Drift: "Frontier Firm"

But marketing doesn't like definitions that reserve a word for ten players. And Microsoft sensed the opening.

Since 2025, Microsoft has actively pushed the term Frontier Firm in its Work Trend Index, its Ignite keynotes, and its Copilot campaigns. The definition is radically different: a Frontier Firm isn't defined by its size or its industry, but by its mindset and execution — AI-first adoption, embedding agents in processes, observability, and so on.

OpenAI joined the movement in May 2026 with its OpenAI Frontier platform, which designates… an enterprise product, not a foundation model.

In short, the word has been broadened by tool vendors to qualify the customers who adopt their tools. It's a smart positioning move: it flatters the buyer, it makes the label desirable, it creates a category you can reach — not just admire.

It's also what makes the word ambiguous.

Why It's Tempting for an SMB

Let's be honest: calling yourself frontier feels good. It signals that you're not in the average. That AI isn't some vague R&D initiative but the backbone of the business model. That the organization is at the front of the train, not at the back.

For a Quebec SMB — particularly in an ecosystem where public and institutional buyers are sometimes several cycles behind the state of the art — it's a powerful communication shortcut.

And there is a kernel of truth. An organization that has genuinely architected its product around foundation models, that orchestrates multi-step agents, that does intelligent routing between models based on cost and task, that measures its token consumption: that organization is doing something qualitatively different from a company that bolted a chatbot onto its FAQ.

The word captures that difference. The problem is that it captures it just as well for the company that only bolted on the chatbot.

Why It's a Trap

In front of an uninformed buyer, saying "frontier" works. In front of an informed buyer — a VP of technology at a major integrator, an investment committee, an acquisition partner — the word triggers a different reaction: prove it.

And from there, two scenarios:

1. You can detail your stack: models used, agent architecture, token governance, observed metrics, the technical debt you've consciously taken on. The label holds.
2. You can't. The label becomes a red flag. Not just "they're exaggerating", but "they don't know the difference between training and consuming". Credibility lost.

A word that's too loaded for what you can back up doesn't strengthen a message. It weakens it.

The trap is especially nasty in a positioning exercise in front of a sophisticated buyer. Someone who has already heard pitch after pitch from "frontier" companies has learned to test the word in the first thirty seconds.

The Real Question: What Are You, Really?

Rather than asking whether you're frontier, I find it more useful to ask two separate questions.

First question: do you train models?

If the answer is no — and it is, for just about every organization that will read this post — then you're not a frontier AI company in the technical sense. Period. That's not a critique, it's a fact. A law firm isn't Anthropic; Anthropic isn't a law firm. Nobody needs to feel ashamed of the distinction.

Second question: do you exploit the best available models, in an architecture that genuinely takes advantage of them?

If the answer is yes, you're what I'd call — without dressing it up — an applied AI-native organization. That's a very defensible status, and it doesn't require hijacking a word.

The Right Vocabulary

There are more precise, more defensible terms that often do a better job:

• AI-native company: a product built around AI from day one, not AI bolted on later.
• Applied AI company: applies existing models to a specific domain with real vertical expertise.
• Agentic AI company: agent-oriented architecture, multi-step orchestration.
• Frontier-adjacent: actively exploits frontier models, sometimes with partnerships or early access, without claiming to train them.

For most Quebec SMBs doing real AI, applied AI or AI-native is both more honest and more differentiating. Frontier Firm in the Microsoft sense can be used, but it also signals that you're speaking in a vendor's vocabulary. That's not neutral.

Concretely, How to Position

If you're revising your site, your pitch, your teaser, or your sales materials, here are a few guardrails.

Don't use a word you couldn't defend in front of the most demanding customer in your market. If your VP of engineering isn't comfortable holding the line, the word is too big.

Be specific about what you do, not about the category you slot into. "We orchestrate Opus, Sonnet, and Haiku agents based on task complexity, with per-project token governance" is more powerful than "we're a frontier firm". The first one tells. The second one claims.

Anchor your positioning in your vertical. "AI-native software company applied to the Quebec education sector" is more differentiating than any generic label. It's also harder to imitate, because it's true.

Beware of words that change meaning every quarter. Frontier is in that category today. AI-first was. Cloud-native was. Rule of thumb: if a word is used by everyone, it no longer differentiates you from anyone.

Keep "frontier" for when you talk about the models, not about yourself. "We exploit frontier models in an agentic architecture" is a fair, defensible sentence that does the work. "We're a frontier firm" is a sentence that demands to be proven at every interaction.

What's Next?

AI vocabulary will keep moving fast. Frontier will probably dilute further — as more and more players claim it, the word will end up signaling nothing at all, the way "digital transformation" ended up signaling nothing by the late 2010s.

The organizations that will stand out won't be the ones who collected the right keywords. They'll be the ones who collected the right proof.

---

Next time you're tempted to write "frontier" on a slide, ask yourself: am I describing a technical reality, or am I decorating a slide? If it's the second, drop the word. You'll be more credible with what's left.